A multispectral remote sensing image super-resolution method and system with distortion perception routing and parameter sharing mechanism

Abstract

The application relates to the fields of image processing and remote sensing image technology, in particular to a multispectral remote sensing image super-resolution method and system with distortion perception routing and parameter sharing mechanism. The method first encodes shallow features of a low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial positions; then extracts frequency domain and spatial domain distortion clue features from the features and encodes the features into distortion feature tokens; completes adaptive routing distribution of the image token features to different expert networks through a gating mechanism and a capacity constraint; performs nonlinear feature reconstruction through a hybrid expert network with cross-spectral band shared backbone parameters and band-level lightweight parameter adaptation; aggregates the output results after weighting, fuses residual errors of the shallow token features, forms high-resolution feature representation, and finally converts into a target high-resolution multispectral remote sensing image, thereby improving the reconstruction effect and optimizing the utilization of computing resources.

Description

Technical Field

[0001] This application relates to the fields of image processing and remote sensing imaging technology, and in particular to a multispectral remote sensing image super-resolution method and system with distortion-aware routing and parameter sharing mechanism. Background Technology

[0002] Remote Sensing Image Super-Resolution (RSISR) aims to reconstruct high-resolution images with rich high-frequency details and consistent structure across spectral bands from low-resolution multispectral images under strict memory and computational constraints. However, in real-world remote sensing scenarios, different bands are affected by various forms of distortion, including edge blurring, loss of texture details, and sensor / band-related random noise. These distortions exhibit heterogeneity in spatial distribution and across spectral bands, often resulting in poor performance with a single, uniform super-resolution restoration strategy. Existing methods struggle to simultaneously restore high-frequency details and maintain cross-band structural consistency, easily leading to situations where details in some bands are enhanced while the overall structure is mismatched.

[0003] Current mainstream super-resolution algorithms (such as models based on convolutional neural networks and Transformers) have achieved excellent performance in super-resolution of natural images, but they face new challenges in remote sensing multispectral images. On the one hand, these models typically only implicitly utilize prior knowledge in the frequency and spatial domains of the image, lacking explicit mechanisms to locate and process specific types of distortion regions. While these models are capable of synthesizing detailed textures, they cannot perform targeted processing based on different distortion types. On the other hand, simply training independent super-resolution models for each spectral band, or directly using large-capacity hybrid expert models to improve performance, will lead to parameter inflation and a sharp increase in memory / computational overhead. Traditional multi-branch or expert models, without effective parameter sharing mechanisms, will result in unacceptable resource consumption due to specialization for each band. Especially in multispectral scenarios, the consistency of outputs across different bands is difficult to guarantee: processing each band separately may lead to misalignment of geometric details or artifacts (such as checkerboard artifacts or spectral aliasing). Therefore, existing technologies lack an efficient super-resolution method that can both explicitly identify and process heterogeneous distortions and maintain cross-band structural consistency with limited resources. Summary of the Invention

[0004] The purpose of this application is to overcome the above-mentioned defects of the prior art and provide a multispectral remote sensing image super-resolution method and system with distortion-aware routing and parameter sharing mechanism, which can reconstruct high-resolution images with consistent details across bands based on low-resolution multispectral images.

[0005] To address the aforementioned technical problems, the technical solution provided in this application offers a multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanisms, comprising: Step 1: Use the shallow feature extraction module to encode the input low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial location; Step 2: Using the distortion-aware feature extractor module, extract distortion cue features reflecting high-frequency missingness, edge degradation, and noise level in the local area from the image token features in the frequency domain and spatial domain, respectively, and explicitly encode the distortion cue features into distortion feature tokens that correspond one-to-one with the image token features. Step 3: Input the distortion feature token and the corresponding image token feature into the distortion perception expert selection routing module. Calculate the affinity relationship between the image token feature and multiple expert networks through a gating mechanism, and perform expert selection routing under capacity constraints. The capacity constraint is used to limit the number of distortion feature tokens assigned to each expert network to be processed to not exceed a preset upper limit, so that image token features with different distortion features are adaptively assigned to different expert networks for processing. Step 4: Through the parameter-sharing hybrid expert network module, nonlinear feature reconstruction is performed on the image token features assigned by the distortion-aware expert selection routing module. Multiple expert networks share expert backbone network parameters across different spectral bands, and band-level lightweight parameter adaptation is introduced for different spectral bands. Step 5: Perform weighted aggregation on the image token feature outputs from one or more expert networks. The weights of the weighted aggregation are obtained by normalizing the affinity scores output by the gating mechanism, and are fused with the image token features obtained in Step 1 through residual method to form a layer-by-layer enhanced high-resolution feature representation. Step 6: The layer-by-layer enhanced high-resolution feature representation is converted into a target high-resolution multispectral remote sensing image through the reconstruction module.

[0006] As an improvement to the aforementioned technical features, the distortion-aware feature extractor module includes a frequency domain branch and a spatial distortion branch. The frequency domain branch is used to project the image token features onto a preset fixed-frequency basis function to obtain spectral response features of different frequency components. The spatial distortion branch is used to locally reconstruct the image token features and extract the gradient edge strength features, high-frequency residual features, and noise level features of the corresponding region. The spectral response features, gradient edge strength features, high-frequency residual features, and noise level features are then fused to generate the distortion feature token.

[0007] As an improvement to the aforementioned technical features, the distortion-aware expert selection routing module includes a gating network. The gating network takes the distortion feature token and the corresponding image token feature as input, and outputs an affinity score for each image token feature relative to multiple expert networks. This score characterizes the degree to which each expert network adapts to the processing of the image token feature. The affinity score output by the gating network is subjected to temperature scaling and / or normalization to adjust the difference in affinity scores between different expert networks, thereby improving the discrimination capability and stability of the distortion-aware expert selection routing module.

[0008] As an improvement to the aforementioned technical features, the distortion-aware expert selection routing module employs a capacity-limited expert claiming strategy for route allocation. Under this strategy, for each expert network, only the K distortion feature tokens with the highest affinity scores in the corresponding dimension of the expert network are allowed to claim from all distortion feature tokens. Furthermore, the number of distortion feature tokens claimed by the expert network does not exceed a preset capacity limit, thereby limiting the computational load of a single expert network during the routing phase. Here, K is a preset capacity parameter for the expert network.

[0009] As an improvement to the above-mentioned technical features, when multiple expert networks attempt to claim the same distortion feature token, the distortion-aware expert selection routing module performs conflict handling based on the following rules: comparing the affinity scores of the distortion feature tokens in each expert network attempting to claim it; allocating the distortion feature token to the expert network with the largest affinity score and whose current number of claimed distortion feature tokens has not reached the preset capacity limit; the successfully allocated distortion feature token will no longer participate in the claiming process of other expert networks.

[0010] As an improvement to the above-mentioned technical features, a distortion feature token corresponding one-to-one with the image token feature is determined to be unclaimed by any expert network when any of the following conditions are met: the distortion feature token has not entered the claim set of the top K distortion feature tokens sorted from high to low based on affinity scores in all expert networks; the distortion feature token has entered the claim set of the top K distortion feature tokens sorted from high to low based on affinity scores in at least one expert network, but the at least one expert network has reached its capacity limit; in the case that the distortion feature token has not been claimed by any expert network, the distortion perception expert selection routing module allocates the distortion feature token to an alternate processing path for feature processing to ensure that all distortion feature tokens are covered.

[0011] As an improvement to the aforementioned technical features, during the training process, the distortion-aware expert selection routing module introduces a load balancing constraint. This load balancing constraint is used to limit the distribution of the number of distortion feature tokens received by different expert networks within the training period to meet a preset deviation range, thereby preventing expert networks from being left unassigned processing tasks for extended periods. The distortion-aware expert selection routing module achieves a controllable trade-off between super-resolution reconstruction quality and computational resource consumption by adjusting the preset capacity parameters of the expert networks.

[0012] As an improvement to the aforementioned technical features, the parameter-sharing hybrid expert network module includes multiple expert sub-networks configured in parallel. Each expert sub-network is used to perform nonlinear feature mapping operations on the assigned image token features. Each expert sub-network includes at least two fully connected mapping layers, and a nonlinear activation function is set between adjacent fully connected mapping layers to perform nonlinear feature transformation and detail generation operations on the input image token features. The multiple expert sub-networks share the same backbone network structure and corresponding backbone parameters across different spectral bands.

[0013] As an improvement to the above-mentioned technical features, a low-rank adapter structure is introduced in at least part of the fully connected mapping layer of the expert subnetwork to achieve band-differentiated adjustment; the low-rank adapter structure includes a set of trainable weight matrices, which perform band-level incremental adjustment on the shared backbone parameters through low-rank matrix decomposition to achieve cross-spectral parameter sharing and band-specific feature adaptation.

[0014] To achieve another objective of the present invention, the present invention also provides a multispectral remote sensing image super-resolution system with distortion-aware routing and parameter sharing mechanisms, comprising: The shallow feature extraction module is used to encode the features of the input low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial location. The distortion-aware feature extractor module is used to extract distortion cue features reflecting high-frequency missingness, edge degradation and noise level in local areas from the image token features in the frequency domain and spatial domain, respectively, and to explicitly encode the distortion cue features into distortion feature tokens that correspond one-to-one with the image token features. The distortion-aware expert selection routing module is used to receive the distortion feature token and the image token feature corresponding to the distortion feature token, calculate the affinity relationship between the image token feature and multiple expert networks through a gating mechanism, and perform expert selection routing under capacity constraints. The capacity constraints are used to limit the number of distortion feature tokens assigned to each expert network to be processed to not exceed a preset upper limit, so that image token features with different distortion features are adaptively assigned to different expert networks for processing. A parameter-sharing hybrid expert network module is used to perform nonlinear feature reconstruction on image token features assigned by the distortion-aware expert selection routing module, wherein multiple expert networks share expert backbone network parameters across different spectral bands and introduce band-level lightweight parameter adaptation for different spectral bands. An aggregation module is used to perform weighted aggregation of image token feature outputs from one or more expert networks. The weights of the aggregation are obtained by normalizing the affinity scores output by the gating mechanism, and are fused with the image token features obtained by the shallow feature extraction module through residual methods to form a layer-by-layer enhanced high-resolution feature representation; and The reconstruction module is used to convert the layer-by-layer enhanced high-resolution feature representation into a target high-resolution multispectral remote sensing image.

[0015] The advantage of this application lies in its core design concept of explicitly detecting image distortion features and adaptively allocating computing resources. It constructs a complete process from distortion cue extraction to expert specialization processing and then to gradual fine reconstruction through three core modules: a distortion-aware feature extractor module, a distortion-aware expert selection routing module, and a parameter-sharing hybrid expert network module. This approach balances reconstruction quality and computing resource utilization efficiency, demonstrating significant advantages. First, the distortion-aware feature extractor module explicitly extracts distortion cues such as high-frequency missing values, edge degradation, and noise levels from images, overcoming the limitations of implicit distortion learning in traditional methods. This provides precise distortion feature data for subsequent differentiated processing, enabling highly targeted super-resolution reconstruction. Second, the distortion-aware expert selection routing module adaptively allocates computing resources on demand based on the extracted distortion features, matching image regions with different distortion features to suitable expert networks for processing. This ensures that computing power is precisely applied to the corresponding distortion regions, significantly improving the efficiency of computing resource utilization. Third, the parameter-sharing hybrid expert network module, based on a cross-spectral band parameter-sharing design, effectively avoids parameter inflation while meeting the needs of different spectral band feature adaptation, significantly reducing the model's computing power and memory overhead. Fourth, the three core modules work together to form an integrated reconstruction process, progressively refining the image. By efficiently utilizing parameters and computing power, this significantly improves the super-resolution reconstruction effect of multispectral remote sensing images, achieving a dual optimization of reconstruction quality and computational efficiency. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall architecture of the method of the present invention; Figure 2 This is a schematic diagram of the Distortion-Aware Feature Extractor (DAFE) module. Figure 3 This is a schematic diagram of the Distortion Awareness Expert Routing (DAEC) framework; Figure 4This is a schematic diagram of the parameter-sharing hybrid expert network PS-MoE architecture; Figure 5 This is a multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism provided in Embodiment 1 of the present invention. Detailed Implementation

[0017] The technical solutions provided in this application are further illustrated below with reference to the embodiments.

[0018] Example 1 This embodiment provides a multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanisms. This method introduces three core modules: distortion feature extraction, expert route allocation, and cross-band parameter sharing, forming a unified process from distortion cue extraction to expert specialization and then to progressively refined reconstruction. The overall design idea of ​​this invention is to explicitly detect distortion features in the image within the model and adaptively allocate computational resources (experts) based on these features, thereby improving the super-resolution reconstruction effect while efficiently utilizing parameters and computing power. The main modules and innovations are as follows: The DAFE (Distortion-Aware Feature Extractor) module extracts distortion cue features in both the frequency and spatial domains from the input image. A lightweight feature extractor, DAFE extracts high-frequency texture loss information by combining fixed frequency basis functions with learnable residuals, and captures spatial distortion features using edge detection (e.g., the Sobel operator) and noise estimation. Specifically, DAFE comprises two branches: the frequency branch projects image features onto a predefined discrete cosine transform (DCT) basis to obtain the coefficients of each frequency component, and applies a learnable gain factor to adapt to the spectral bias of a specific image; the spatial branch reconstructs the features back into image patches, calculates their gradient edge strength and high-frequency residuals, and estimates the noise component by suppressing major edge components. The features obtained from the two branches (frequency vectors and scalars representing edge and noise levels) are concatenated and normalized to generate "distortion feature tokens." These distortion tokens are explicit descriptions of the degree of distortion in different regions of the image and will be used by the routing module later. The DAFE module has a simple and efficient structure, introducing only a small number of matrix operations and convolution operations, which has little impact on the overall model overhead, but provides sensitive awareness of the location and type of distortion.

[0019] The DAEC module (Distortion-Aware Expert Selection Router) is a mechanism for dynamically routing subsequent expert networks based on distortion features. The DAEC module includes a gating network that takes distortion feature tokens from the DAFE (Distortion Awareness Provider Interface) and basic image features as input, and outputs an affinity score for each expert network. The router employs a capacity-constrained expert selection mechanism: each expert network can only receive a limited number of image feature "tokens," i.e., a maximum number of tokens each expert can process (capacity threshold) is set. During runtime, the router selects the Top-K tokens (where K is the expert's capacity) from all tokens based on the affinity score, ensuring that each expert only processes the batch of regions with the highest score, thus achieving on-demand computation. If multiple experts want to process the same token or some tokens are not claimed by any expert, the DAEC module also defines conflict resolution and fallback strategies: unclaimed tokens will be processed through a default lightweight channel, ensuring that no input region is ignored. Simultaneously, the router limits each token to being processed by a maximum of a few different experts simultaneously (limiting the fan-out of tokens) to control computational redundancy. By introducing capacity constraints and load balancing strategies, DAEC achieves predictable memory and latency control, enabling the model to effectively balance image quality and computational resources when processing high-resolution multi-band data. In short, the DAEC module adaptively assigns different regions of the image to corresponding experts based on the distortion information provided by DAFE, ensuring that "severely distorted regions are handled by stronger experts, while simpler regions are handled with fewer resources," thereby improving processing efficiency and effectiveness.

[0020] The PS-MoE module (Parameter-Sharing Hybrid Expert Network) is a hybrid expert network used for specialized reconstruction of image features, featuring parameter sharing across spectral bands. The PS-MoE module comprises multiple parallel expert sub-networks, each typically consisting of two fully connected neural network (MLP) layers for nonlinear transformation and detail reconstruction of input features. In this invention, these expert networks employ a parameter-sharing strategy, using a unified set of expert weights across all bands instead of storing a separate set of parameters for each band. To account for statistical differences between images in different bands, each expert network introduces a low-rank adapter (LoRA) module in its key layers for band-specific fine-tuning. Specifically, in the two-layer MLP mapping of the expert network, a set of low-rank incremental weights (with a parameter size much smaller than the original network) is added for each spectral band. These LoRA modules endow the experts with fine-tuning capabilities across bands, allowing them to adjust for specific noise distributions or detail characteristics within that band, while the underlying main weights remain shared across all bands. This design significantly reduces redundant parameters and storage overhead, achieving near-instantaneous results for training experts for each band, while adding only a minimal number of parameters. Furthermore, by sharing an expert backbone network, this invention ensures the consistency of structural and geometric details in the reconstruction results across spectral bands, preventing cross-band misalignment or artifacts. In the PS-MoE module, each token from the DAEC routing selection is sent to the corresponding expert for processing. If a token is claimed by multiple experts, their outputs are aggregated according to weights. Through the layer-by-layer processing of the PS-MoE module, the image is gradually reconstructed from coarse structure to fine texture, ensuring high fidelity while effectively controlling the number of model parameters and computational load.

[0021] In summary, this invention proposes a multispectral super-resolution framework that prioritizes detection followed by restoration and allocates resources on demand. First, explicit distortion cue features are extracted using the DAFE module. Then, the DAEC module performs capacity-aware expert routing decisions. Finally, the PS-MoE module shares expert parameters across bands for efficient detail restoration. This scheme fully utilizes prior knowledge in both the frequency and spatial domains, allocates different processing capabilities to different distortion types, and avoids the overhead of band-specific models through parameter sharing, achieving an ideal balance between image quality and computational efficiency. Extensive experimental results also demonstrate that this method significantly reduces floating-point computation and the number of parameters while achieving super-resolution reconstruction performance that meets or exceeds existing state-of-the-art Transformer benchmark models.

[0022] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but the present invention is not limited to the embodiments described herein.

[0023] Figure 1This is a schematic diagram of the overall architecture of the method of this invention. This invention is designed based on the Swin2SR model backbone and includes: (1) Shallow feature extractor, used to extract preliminary features from the original low-resolution image; (2) A deep backbone network composed of multiple residual Swing Transformer blocks stacked together, in which the distortion-aware routing and expert module of the present invention is integrated to refine image features step by step; (3) Reconstruction module, which reconstructs the processed features back into a high-resolution multispectral image.

[0024] The fully connected layer (FFN) in the original Transformer backbone is replaced by the distortion-aware sparse expert module of this invention, thereby introducing distortion-aware routing and parameter sharing expert mechanisms. The PS-MoE module achieves a layer-by-layer refinement process of prioritizing shallow-layer structure reconstruction and deep-layer texture optimization by adding a band-level LoRA adapter to the shared MLP expert backbone.

[0025] Figure 2 This is a schematic diagram of the Distortion-Aware Feature Extractor (DAFE) module. The DAFE module consists of two parts: a frequency branch (left) and a spatial distortion branch (right). The left-hand frequency branch projects the image token onto a fixed set of DCT basis functions (with learnable gain parameters) to explicitly obtain the coefficients of different frequency components. The right-hand distortion branch reconstructs the token into small image patches, calculates its Sobel gradient edge strength and high-frequency residual signal, and extracts the noise surrogate by comparing the high-frequency residual with the edge strength. The spectral vector obtained from the frequency branch and the edge and noise scalars obtained from the spatial branch are finally concatenated, and after layer normalization and linear transformation, the distortion feature token used for routing is output.

[0026] Figure 3 This is a schematic diagram of the Distortion-Aware Expert Selection Routing (DAEC) framework. The distortion features extracted by the DAFE module are first fused with regular image features and input into a gating network to generate an affinity score for each expert. Based on the score, the router allows each expert to "claim" the Top-K tokens they are most confident in processing (where K is the expert's capacity limit), implementing a capacity-constrained expert selection strategy. Tokens not claimed by any expert are sent to a backup lightweight processing path to ensure all image regions are processed. The scheduler distributes each token to the corresponding expert for processing based on the routing decision, while the aggregator fuses the outputs of all experts through residual connections to obtain the output features of the current layer.

[0027] Figure 4This is a schematic diagram of a parameter-sharing hybrid expert network (PS-MoE) architecture. This module contains multiple parallel expert MLP networks (gray areas), with the backbone parameters of all experts shared across different bands. Each expert's two-layer fully connected map is supplemented with a band-specific LoRA adapter (represented by colored modules) to provide fine-tuning of parameters for each band without duplicating the entire set of expert parameters. The DAEC router delivers a sparsely gated signal to the selected expert, who transforms the token features and outputs the result. The aggregator performs a weighted summation of the outputs of multiple experts for a single token and fuses the expert outputs with the original features through residual connections. This architecture achieves near-individual expert configuration for each band while significantly reducing parameter and memory overhead.

[0028] like Figure 1 As shown, the method of this invention employs an improved Transformer architecture as the backbone network to effectively integrate distortion-aware routing and parameter-sharing expert mechanisms, achieving super-resolution reconstruction of multispectral images. The entire model can be divided into three stages: input feature extraction, layer-by-layer distortion-aware recovery, and output reconstruction. Figure 5 As shown, the multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism provided in this embodiment first processes the input low-resolution multispectral image through a shallow feature extraction module, converting the image of each band into an initial feature representation (e.g., mapping image pixels to a high-dimensional feature space through convolution). Next, the features are fed into a deep backbone network for progressive reconstruction and enhancement. Each layer of the backbone network embeds the DAFE, DAEC, and PS-MoE modules of this invention to specifically enhance the image features. The main steps of the method of this invention are illustrated below using a typical implementation process as an example: Step 1: Distortion-Aware Feature Extraction (DAFE). At the beginning of each backbone layer, the corresponding distortion-aware feature token is calculated for the image features input to the current layer. Specifically, the feature vector at each spatial location is processed using the DAFE module. On one hand, the feature is projected onto a set of predefined DCT bases through the frequency branch to obtain a vector representing the intensity of each frequency component in that region. Since a fixed DCT base may not perfectly fit all images, we introduce a learnable residual linear layer (initialized to zero weights) after projection, allowing the model to automatically learn the spectral biases relevant to the specific task during training. This ensures that the output frequency vector contains both interpretable standard frequency components and allows for adjustments for the current task. On the other hand, the feature at that location is mapped back to the original image space through the spatial distortion branch (i.e., a small image patch is reconstructed), and its gradient edge map and smoothed residual map are calculated: First, the Sobel operator is applied to extract the edge intensity of the image to obtain the edge magnitude map of the patch; then, mean filtering or convolution smoothing is applied to the patch image, and the smoothing result is subtracted to obtain the high-frequency residual map. Next, the high-frequency residual is subtracted from the edge amplitude and a positive value is taken (e.g., ReLU activation) to obtain a noise proxy map mainly containing random noise. Then, the edge amplitude map of this patch is averaged to obtain an edge intensity scalar, and the noise proxy map is averaged to obtain a noise intensity scalar, used to summarize the edge sharpness and noise level near that location. Finally, the DAFE module concatenates the frequency vector with the corresponding edge and noise scalars, performs layer normalization and linear transformation, and generates a low-dimensional distortion feature token (e.g., ...). Figure 2 (As shown). This token explicitly describes the degree of texture detail loss and noise level in the current local image, serving as the basis for subsequent routing decisions. It should be noted that if, in some cases, real noise or edge information cannot be directly obtained (e.g., unregistered input or missing data for a certain band), the DAFE module can degenerate to provide features based solely on frequency components, or use a default approach to approximate edge / noise proxies to ensure the robustness of the entire process. The DAFE extraction process has a very low computational cost, involving only one matrix multiplication and a small number of convolution operations, and uses low-precision computations such as FP16 / BF16 (FP32 is used for edge and normalization parts to ensure numerical stability), making it ideal for real-time execution on resource-constrained devices.

[0029] Additional information for Step 1: Dimensions and basic symbols: The input token feature matrix represents the set of token-level features of the image after shallow / backbone extraction. ; The number of tokens corresponds to the spatial location / number of image tiles; For single-token feature dimensions; For the real number field, Indicates by A matrix space consisting of n real elements; This represents the number of frequency components (DCT modes) and the corresponding number of channels in the frequency branch output. bs is the batch size, which is the number of samples processed in parallel at one time. The height of the reconstructed image patch (or feature map) in the spatial domain; The width of the reconstructed image patch (or feature map) in the spatial domain; T is the matrix transpose operator; Indicates the vector A diagonal matrix formed by placing the elements on the diagonal.

[0030] Frequency branch: A fixed frequency basis matrix (e.g., a fixed DCT basis) is used to project input features onto the frequency basis space. Its dimensions are given by the formula. Compatible (i.e.) , equivalent to ); This is a learnable frequency gain vector used for modal scaling of the projected response for each frequency mode, and... Compatible.

[0031] The "fixed frequency response" matrix after projection onto the frequency basis:

[0032] Each row corresponds to a token. The response on each frequency mode.

[0033] This is a learnable residual projection matrix (residual head parameter) used for learnable compensation of the frequency projection, and... Compatible (i.e.) ).

[0034] The final output frequency characteristic matrix of the frequency branch:

[0035] This represents the frequency domain distortion cue representation for each token.

[0036] Spatial branch: To reconstruct (or intensity map) the weight vector / matrix, used to map token features to scalar strength, and... Compatible.

[0037] For token-level strength vectors (one scalar strength per token):

[0038] The reshape() function represents a reshaping operation, used to restore a sequence of tokens to a two-dimensional spatial arrangement.

[0039] I represents the reshaped image patch / feature map, indicating: The channel number is 1, which indicates the intensity map used for distortion analysis.

[0040] Edge strength: This is the horizontal convolution kernel of the Sobel operator.

[0041] This is the convolution kernel of the Sobel operator in the vertical direction; This represents the convolution operator; For horizontal gradient response:

[0042] Vertical gradient response:

[0043] It is a numerical stability constant (a non-negative small scalar) used to avoid gradient energy being zero or subsequent computations being unstable.

[0044] For gradient energy map (edge ​​strength surrogate):

[0045] High-frequency residuals and noise proxy: AvgPool(•) represents the average pooling / smoothing operator, used to generate a smoothed version.

[0046] High-frequency residual amplitude plot:

[0047] To modify the activation function of the linear unit, the non-negative part is selected. Enter the token index.

[0048] Noise proxy diagram :

[0049] Vectorization, scaling, and scalars: As a vectorization operator, the space tensor is flattened in token order. The column vector representation of N represents the number of tokens, i.e., the number of elements corresponding to the spatial feature map after it is flattened.

[0050] This is the edge intensity scaling factor (learnable or preset scalar), used to adjust the weight of edge cues in the distortion features.

[0051] This is a noise intensity scaling factor (learnable or preset scalar) used to adjust the weight of noise cues in the distortion features.

[0052] For edge strength scalars (token level):

[0053] For noise intensity scalar (token level):

[0054] splicing and projection: This represents the concatenation operator based on feature dimensions.

[0055] LN(•) represents the layer normalization operation.

[0056] For distorted clue fusion feature matrix: It is a linear projection function used to map fused features to a unified dimensional space.

[0057] This represents the dimension of the distortion feature after projection (the dimension of the distortion token).

[0058] For the distortion feature token (DAFE output):

[0059] Routing gating: The base image token features (backbone features used for routing and subsequent expert processing) have dimensions compatible with the inputs of subsequent gating networks.

[0060] To analyze basic image features The result of layer normalization, For basic features The applied linear transformation / feature mapping result ("LF" stands for Linear / LatentFeature Transformation Module) is used to fuse with the distorted features.

[0061] For use in Linear weight matrix mapped to the gated input space.

[0062] || represents the feature concatenation operator (AND) Equivalence means concatenating along the feature dimension.

[0063] For the fusion input features of the routing gated network:

[0064] This feature is used in subsequent DAEC gating network calculations of affinity scores.

[0065] Step 2: Expert Routing Decision (DAEC). For each backbone layer, the distortion feature tokens (representing the distortion at each spatial location) generated by the DAFE module above are input into the Distortion-Aware Expert Selection Router (DAEC), used together with the basic features of the original image to determine which expert networks will handle different locations within that layer. For example... Figure 3As shown, the DAEC module first calculates a set of expert affinity scores for each token using a gating network. The gating network can be understood as several fully connected neurons that output a score for each expert; these scores reflect the expert's proficiency in processing that token. To improve routing discriminative power, the gating network output is typically temperature-scaled or normalized to adjust the score differences. Next, DAEC executes a capacity-constrained Top-K selection strategy based on the score matrix: each expert only selects the K tokens with the highest scores for processing (K is a pre-set capacity for that expert). For example, if K=10 and an expert has high scores for several image regions, it only takes the 10 regions with the highest scores, ignoring the others even if they have high scores. This process is equivalent to each expert selecting the image patches they are most confident in processing, ensuring that each expert's processing load does not exceed the designed capacity, thus achieving a controllable allocation of computational overhead for each expert. When multiple experts attempt to take the same token, conflict resolution is performed based on scores and capacity limitations: the expert with the highest score and remaining capacity takes priority, and the token becomes unavailable to other experts. If a token is not claimed by any expert (possibly because all experts' capacities are full and the token's ranking among all experts is not in the Top-K), then the token will go through the backup path. The backup path can be a simplified channel (e.g., a shared small convolution or skipping complex processing altogether) to ensure that the token is still processed and not completely ignored. Through this mechanism, DAEC dynamically assigns different parts of the image to different experts, allowing each expert to focus on processing the distortion types they are good at. For example, some experts may focus on reconstructing high-frequency detail textures, while others handle blurred and smooth areas, each taking advantage of their strengths. It is worth mentioning that DAEC can incorporate regularization terms such as load balancing constraints during training to encourage all experts to be utilized without collapse (i.e., a few experts continuously process the majority of tokens while others remain idle). Furthermore, by adjusting the expert capacity K value, the system provides a trade-off between quality and efficiency: a larger capacity allows more tokens to be processed by more refined experts, improving reconstruction quality but increasing computation; a smaller capacity limits the computational input per layer, reducing memory and latency overhead. In engineering implementation, appropriate capacity parameters can be selected based on equipment performance and task requirements to achieve a dynamic balance between performance and resources.

[0066] Step 2 Supplement: DAEC Capacity-Constrained Expert-Choice Routing Algorithm: Enter the relevant symbols: For the input token set:

[0067] in, For the first The basic feature vector corresponding to each image token.

[0068] Enter the total number of tokens.

[0069] The set of distortion feature tokens output by the DAFE module:

[0070] In order to be with the first The distortion feature vector corresponding to each input token.

[0071] For the gathering of experts:

[0072] Each of them An index representing an expert network; This represents the total number of expert networks.

[0073] This is the capacity factor, used to determine the processing capacity of a single expert based on the number of tokens and the number of experts. It is a numerical parameter.

[0074] The constraint factor (fan-out factor) represents the maximum number of experts to whom a single token can be assigned. Constraint: Each token can be received by at most ρ experts (ρ∈{1,2}).

[0075] Fusion input and gating feature definition: The layer normalization operation is represented.

[0076] This represents the feature concatenation operator, which indicates concatenation along the feature dimensions.

[0077] It is a linear mapping weight matrix used to map distorted features to a space compatible with the basic features.

[0078] Symbols related to affinity (score) calculation: is the linear mapping weight matrix of the gated network, used to map fused features to expert affinity scores.

[0079] This is a temperature parameter used to adjust the smoothness of the affinity fraction distribution.

[0080] For the first The token pair The expert's approachability (rating).

[0081] The input features are used as input features to the gating network, and are subsequently used to calculate the affinity score of the token relative to each expert network.

[0082]

[0083] Capacity settings and expert application: The capacity limit for a single expert, i.e., the maximum number of tokens each expert can receive, is defined as follows:

[0084] This is the floor operator.

[0085] This indicates selecting the largest value from the input set. Operators with 1 element.

[0086] Indicates being the first A set of token indexes claimed by experts:

[0087] Conflict handling and token allocation: The maximum number of experts that can be assigned to a single token, satisfying the following conditions: .

[0088] This indicates selecting the largest value from the input set. Operators with 1 element.

[0089] For the actual allocation processing A collection of experts for each token:

[0090] Weight normalization and expert output aggregation: This is a normalization function used to convert multiple expert scores into a distribution with a weighted sum of 1.

[0091] For the first The expert in handling the first Normalized weights for tokens:

[0092] Indicates the index number of the expert network, set This indicates that from the expert index set {1,…,E}, based on the token... Affinity scores for each expert And a subset obtained by filtering under capacity constraints, whose elements are still expert indexes. The expression " " indicates that the number is The experts involved in handling the first A collection of experts for each token.

[0093] For the first A forward mapping function of an expert network is used to perform nonlinear feature transformation on the input token features. Indicates the first Then, for each input token, the corresponding basic feature vector is: Indicates the first The expert network on the first The basic feature vector of each input token The output feature vector obtained after forward mapping is used for subsequent feature aggregation and super-resolution reconstruction processing.

[0094] For the first The output characteristics of each token after processing and aggregation by an expert network.

[0095] Expert aggregation output (normal path): when hour:

[0096] Alternate path related symbols: This is a lightweight processing function for the fallback path, used to perform basic feature processing on input tokens that have not been claimed by any expert network. Indicates the first Then, for each input token, the corresponding basic feature vector is: This indicates that the alternative path processing function is for the first... The basic feature vector of each input token The output features obtained after lightweight feature transformation are used for subsequent fusion or residual connection with the expert output of the normal path in the same feature space.

[0097] Alternative path output: when hour:

[0098] This represents an empty set, used to describe the state of the expert set. The expression " "Indicates participation in the processing of the first The set of experts for a token is non-empty, meaning there exists at least one expert who processes the token; this is the semantic condition for set determination.

[0099] Note: Alternate paths can use lightweight shared MLP or Top-1 fallback routes to ensure coverage and stability; and achieve predictable memory / latency without losing tokens.

[0100] Residual output:

[0101] For the first The output feature vector of each token in the current network layer is defined as follows: ; in: Indicates the first The basic feature vector corresponding to each input token; Indicates the first The output characteristics of each token after being processed and weighted by an expert network or alternative path; This represents the token feature after residual connection and layer normalization, which is used as input to subsequent network layers.

[0102] Step 3: Expert Network Processing (PS-MoE). After DAEC determines which experts will process each token, the next step is for a parameter-sharing expert hybrid network (PS-MoE) to perform the actual super-resolution reconstruction of the token features. For example... Figure 4 As shown, PS-MoE consists of multiple parallel expert subnetworks. Each expert is typically implemented as a two-layer fully connected network (MLP) with nonlinear activation, used for nonlinear transformation and detail generation of input features. In traditional methods, equipping each spectral band with a separate set of experts would lead to a significant increase in the number of parameters and memory usage. To avoid this problem, the PS-MoE module of this invention adopts a cross-band parameter sharing scheme: all bands share the same expert backbone with the same structure and parameters, and low-rank adapters (LoRAs) are introduced only where differentiation is needed for adjustment. Specifically, each expert's two-layer fully connected mapping adds a small set of weight matrices (i.e., LoRA modules) at both the input and output projections. These weights are activated only on the corresponding band to provide fine-tuning for that band. For example, if the weight dimension of the expert MLP is m×n, the LoRA adapter can be decomposed into small matrices of m×r and r×n (r... The weights are represented in the form of m, n, as a low-rank update of the original weights. Thus, when processing data for a specific band, the expert's actual weight equals the shared backbone weight plus the LoRA weight offset for that band; while processing another band, a different set of LoRA offsets is applied. Since the value of r can be chosen very small, the increase in parameters and computation for each band is very low, and the impact on overall performance is negligible, but it gives the model the ability to fine-tune by band. During the routing and scheduling phase, DAEC sends each token to one or more selected experts for processing. If a token is assigned to a single expert, the expert is used directly to perform MLP forward computation on the features; if it is processed by multiple experts simultaneously, the system calls these expert networks in parallel, each outputting the transformation result for that token. Subsequently, the aggregator of the PS-MoE module performs a weighted sum of the outputs from multiple experts according to the weights calculated by the previously gated network to obtain the comprehensive output features of that token. Whether using single-expert or multi-expert processing, PS-MoE fuses expert outputs with original features through residual connections to preserve original information and increase stability (similar to the residual structure of the FFN layer in Transformer). As the model depth increases, experts in different layers can gradually assume different responsibilities: shallower experts focus more on reconstructing the basic structure of the image and suppressing artifacts, while deeper experts concentrate on fine texture restoration in high-frequency detail regions marked by DAFE. By stacking multiple "DAEC+PS-MoE" modules layer by layer, the model forms a coarse-to-fine reconstruction process, performing super-resolution enhancement on the image from macroscopic to microscopic levels. This progressive processing approach, combined with distortion-aware routing control, allows computational resources to be allocated hierarchically according to the needs of the image content, maximizing efficiency while ensuring reconstruction quality.

[0103] Step 3 Supplement: LoRA scaling factor related symbols: The scaling hyperparameters for LoRA are used to control the impact of the low-rank adapter on the expert backbone mapping.

[0104] d represents the dimension of the token feature in the expert backbone network. Let r be the rank of the low-rank decomposition of LoRA, satisfying r d, to ensure that the number of parameters and computational overhead introduced by the low-rank adapter are much smaller than those of the shared backbone parameters.

[0105] This represents the number of LoRA layers or a normalization constant (used to maintain scale consistency across different LoRA settings).

[0106] The LoRA scaling factor is given by the following formula:

[0107] This coefficient is used to uniformly adjust the contribution intensity of low-rank incremental branches in forward computation.

[0108] Input and intermediate feature symbols: The base image token feature vector (or a row in the token feature matrix) represents the input features entering the current expert network.

[0109] The hidden features of the shared trunk branch in the first layer represent the nonlinear mapping result calculated only by the shared parameters.

[0110] The intermediate hidden features of the LoRA branch in the first layer represent the incremental adjustments made to the shared trunk mapping by the low-rank adapter.

[0111] This represents the hidden features after the first-level shared branch and the LoRA branch are merged.

[0112] First-level mapping related symbols (shared backbone + LoRA): The weight matrix for the first fully connected layer of the expert backbone network is shared across all spectral bands.

[0113] This is the bias vector of the fully connected mapping of the first layer of the expert backbone network.

[0114] For the first The LoRA input low-rank matrix corresponding to each spectral band at the first layer mapping.

[0115] For the first The LoRA output low-rank matrix corresponding to each spectral band at the first layer mapping is a matrix.

[0116] GELU(•) is the Gaussian error linear unit activation function, used to introduce nonlinear expressive power.

[0117] First-level shared branch output:

[0118] First-level LoRA branch output:

[0119] First layer fusion output:

[0120] Second-level mapping related symbols (shared backbone + LoRA): The weight matrix for the second fully connected layer of the expert backbone network is shared across all spectral bands.

[0121] This is the bias vector for the fully connected mapping of the second layer of the expert backbone network.

[0122] For the first The LoRA input low-rank matrix corresponding to each spectral band at the second layer mapping.

[0123] For the first The corresponding low-rank LoRA output matrix for each spectral band at the second layer mapping.

[0124] Second-level shared branch output:

[0125] Second-level LoRA branch output:

[0126] Expert output and bands: This is a spectral band index used to distinguish the LoRA adaptation parameters corresponding to different spectral channels.

[0127] For the first A forward mapping function for an expert network.

[0128] For the first The expert in handling the first Input features of each spectral band The output result at that time is defined as the superposition of the shared backbone output and the LoRA incremental output:

[0129] Step 4: Reconstructing High-Resolution Image. After multi-level feature processing is completed, the network finally converts the features output by the deep backbone into the target high-resolution image through the reconstruction module. The reconstruction module can adopt conventional designs in existing super-resolution networks, such as upsampling interpolation + convolutional fusion, or sub-pixel convolution (pixel shuffle), to gradually enlarge the feature map to the required resolution. In a typical implementation of this invention, the reconstruction module receives the high-dimensional features output by the backbone network, upsamples the features of each spectral band to the original required size through two sub-pixel convolutional layers, and obtains the reconstructed high-resolution band image through convolution. The reconstruction results of all bands together constitute the final high-resolution multispectral image output. Due to the cross-band parameter sharing and consistent routing mechanism adopted in the backbone network processing, the output images of each band of this invention are highly consistent in geometric details and well aligned, without cross-band artifacts such as misalignment and spikes. At the same time, the details of each band are appropriately enhanced according to its own characteristics. For example, noisy bands are denoised and smoothed by the expert network, while textured bands obtain additional high-frequency detail reconstruction. This unified and flexible processing workflow is applicable to data from various remote sensing imaging satellites, improving the practicality of multispectral remote sensing image super-resolution.

[0130] Example 2 This embodiment provides a multispectral remote sensing image super-resolution system with distortion-aware routing and parameter sharing mechanisms, including: The shallow feature extraction module is used to encode the features of the input low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial location. The distortion-aware feature extractor module is used to extract distortion cue features reflecting high-frequency missingness, edge degradation and noise level in local areas from the image token features in the frequency domain and spatial domain, respectively, and to explicitly encode the distortion cue features into distortion feature tokens that correspond one-to-one with the image token features. The distortion-aware expert selection routing module is used to receive the distortion feature token and the image token feature corresponding to the distortion feature token, calculate the affinity relationship between the image token feature and multiple expert networks through a gating mechanism, and perform expert selection routing under capacity constraints. The capacity constraints are used to limit the number of distortion feature tokens assigned to each expert network to be processed to not exceed a preset upper limit, so that image token features with different distortion features are adaptively assigned to different expert networks for processing. A parameter-sharing hybrid expert network module is used to perform nonlinear feature reconstruction on image token features assigned by the distortion-aware expert selection routing module, wherein multiple expert networks share expert backbone network parameters across different spectral bands and introduce band-level lightweight parameter adaptation for different spectral bands. An aggregation module is used to perform weighted aggregation of image token feature outputs from one or more expert networks. The weights of the aggregation are obtained by normalizing the affinity scores output by the gating mechanism, and are fused with the image token features obtained by the shallow feature extraction module through residual methods to form a layer-by-layer enhanced high-resolution feature representation; and The reconstruction module is used to convert the layer-by-layer enhanced high-resolution feature representation into a target high-resolution multispectral remote sensing image.

[0131] As can be seen from the above implementation methods, this invention combines explicit distortion feature extraction with a sparse expert model, achieving a highly efficient multispectral image super-resolution scheme in engineering. Compared with traditional methods, the method and system of this invention can adaptively allocate computing resources according to image content: more experts are involved in processing complex regions, while simple regions require less intervention, thereby significantly improving the utilization rate of computing resources. While ensuring or improving reconstruction quality, the required number of parameters and computational load are significantly reduced. The modular design (DAFE, DAEC, PS-MoE) provided by this invention is easy to integrate into existing deep learning models, facilitating engineering implementation. For example, in actual deployment, the DAFE module can be pre-embedded in the model as a convolution operator, the Top-K selection of DAEC routes can be implemented through an efficient parallel algorithm, and the parameter sharing mechanism of the PS-MoE expert network can fully utilize the parallel computing capabilities and memory cache of the hardware. A series of experiments show that the method and system provided by this invention achieve leading super-resolution reconstruction results on typical remote sensing datasets, verifying its technical feasibility and superiority. Those skilled in the art can adjust specific parameters (such as the number of experts, capacity K value, LoRA rank value, etc.) based on the above teaching to adapt to different application needs. Without departing from the principles of this invention, various equivalent substitutions and optimizations can be made to the implementation steps and module details. The scope of protection of this invention is defined by the claims.

[0132] 7. Effect Verification and Performance Comparison Under a unified hardware platform (NVIDIA A800-80GB, AMP bfloat16, batch=1) and a unified evaluation protocol, the method and system of this invention achieve a combined effect of "higher quality + lower computational / parameter requirements" on multiple real multispectral / cross-sensor datasets. The following table presents representative results (mean):

[0133] Note: The above comparison results are used to illustrate the technical effects and engineering feasibility of the present invention; the specific values ​​may fluctuate slightly depending on the data partitioning, training configuration and hardware environment, but the overall trend is that while the number of parameters and dynamic computation are reduced, the reconstruction quality (PSNR / SSIM) can reach or exceed the mainstream Transformer benchmark.

[0134] 8 Innovations Compared with Existing Technologies Innovation Point 1: DAFE Distortion-Aware Feature Extraction (Explicitly Routable Distortion Tokens) Compared with existing technologies, common multispectral super-resolution methods rely heavily on implicit learning of networks and lack explicit representations of distortions such as "loss of frequency domain details / noise / edge structure", making it difficult for subsequent enhancement strategies to be targeted according to the type of distortion.

[0135] Distinguishing features: This invention sets up a DAFE module, which adopts a dual-branch structure of "frequency branch + spatial distortion branch": the frequency branch projects the image token onto a fixed DCT basis and adds a learnable gain to obtain the spectral coefficients; the spatial branch reconstructs small blocks and calculates the Sobel edge intensity and high-frequency residual, and extracts the noise proxy quantity from the relationship between the two; finally, the spectral vector and the edge / noise scalar are concatenated, and the distortion feature token is output after normalization and linear transformation for use by the router.

[0136] The resulting technical effects: By making the distorted information explicit and tokenized, subsequent routing can implement differentiated processing for different regions based on the differences in "spectrum / edge / noise", thereby improving the targeting of reconstruction and reducing invalid computation (providing a basis for subsequent sparse expert allocation).

[0137] Innovation Point 2: DAEC Capacity-Constrained Expert Routing (On-Demand Allocation of Controllable Computing Power + Coverage Guarantee) Compared with existing technologies, traditional multi-branch / expert or MoE solutions often suffer from uncontrollable load, "preemption / omission" in some areas, and inference resources becoming unstable as resolution increases.

[0138] Distinguishing features: This invention proposes a DAEC router: the gating network takes DAFE distortion tokens and basic image features as input and outputs affinity scores for each expert; the routing adopts a capacity constraint strategy of "experts claiming Top-K tokens" (capacity threshold per expert) and sets conflict resolution and rollback strategies: those who have not claimed tokens enter the backup lightweight path, while limiting the number of token fanouts to suppress redundancy.

[0139] The resulting technical effects are: while ensuring that all regions are processed, expert load balancing and predictable control of memory / latency are achieved, enabling the model to be deployed stably under high-resolution multi-band data, and achieving on-demand computing with "more computing power allocated to heavily distorted regions and less computing power allocated to simple regions".

[0140] Innovation Point 3: PS-MoE cross-band parameter sharing (shared backbone + lightweight professional LoRA for bands) Compared with existing technologies: While configuring an independent model / independent expert for each band can enhance band specificity, the parameter and memory overhead increases linearly with the number of bands, which is unacceptable for engineering projects; and it is difficult to ensure consistency across bands.

[0141] Distinguishing features: This invention allows expert MLP backbone parameters to be shared across different bands in PS-MoE, and only adds band-specific LoRA adapters to each band at the two mapping layers, achieving band fine-tuning without copying the entire set of expert parameters.

[0142] The resulting technical effects are: achieving near-"individual experts per band" performance while significantly reducing parameter / memory overhead, and making it easier to maintain cross-band structural consistency and controllable training / deployment costs.

[0143] Innovation Point 4: Replace FFN with sparse experts in the Swin2SR backbone (embedding "detection-routing-expert" into the backbone) Compared with existing technologies: The FFN of the existing Transformer super-resolution backbone is usually computationally intensive, making it difficult to make differentiated allocations for different distortion levels in different regions, and it is also difficult to balance quality and computing power.

[0144] Distinguishing features: This invention is based on the Swin2SR design, replacing the FFN in the original Transformer block with a "feature-aware sparse expert module", and integrating the distortion-aware routing and parameter sharing expert mechanism into the residual Swin Transformer stacking backbone to achieve progressively refined reconstruction. The resulting technical effects are: sparse expert processing is directly introduced at the backbone level, achieving a balance between quality improvement and controlled computation; according to the representative comparison results given above, this invention can achieve a comprehensive trend of "higher quality + lower parameters / computation" on multiple datasets.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism, characterized in that, include: Step 1: Use the shallow feature extraction module to encode the input low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial location; Step 2: Using the distortion-aware feature extractor module, extract distortion cue features reflecting high-frequency missingness, edge degradation, and noise level in the local area from the image token features in the frequency domain and spatial domain, respectively, and explicitly encode the distortion cue features into distortion feature tokens that correspond one-to-one with the image token features. Step 3: Input the distortion feature token and the corresponding image token feature into the distortion perception expert selection routing module. Calculate the affinity relationship between the image token feature and multiple expert networks through a gating mechanism, and perform expert selection routing under capacity constraints. The capacity constraint is used to limit the number of distortion feature tokens assigned to each expert network to be processed to not exceed a preset upper limit, so that image token features with different distortion features are adaptively assigned to different expert networks for processing. Step 4: Through the parameter-sharing hybrid expert network module, nonlinear feature reconstruction is performed on the image token features assigned by the distortion-aware expert selection routing module. Multiple expert networks share expert backbone network parameters across different spectral bands, and band-level lightweight parameter adaptation is introduced for different spectral bands. Step 5: Perform weighted aggregation on the image token feature outputs from one or more expert networks. The weights of the weighted aggregation are obtained by normalizing the affinity scores output by the gating mechanism, and are fused with the image token features obtained in Step 1 through residual method to form a layer-by-layer enhanced high-resolution feature representation. Step 6: The layer-by-layer enhanced high-resolution feature representation is converted into a target high-resolution multispectral remote sensing image through the reconstruction module.

2. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 1, characterized in that, The distortion-aware feature extractor module includes a frequency domain branch and a spatial distortion branch, wherein... The frequency domain branch is used to project the image token features onto a preset fixed frequency basis function to obtain the spectral response features of different frequency components. The spatial distortion branch is used to locally reconstruct the image token features and extract the gradient edge intensity features, high-frequency residual features and noise level features of the corresponding region. The distortion feature token is generated by fusing the spectral response features, gradient edge strength features, high-frequency residual features, and noise level features.

3. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 1, characterized in that, The distortion-aware expert routing module includes a gating network. The gating network takes the distortion feature token and the corresponding image token feature as input, and outputs an affinity score for each image token feature relative to multiple expert networks. This score characterizes the degree to which each expert network adapts to the processing of the image token feature. The affinity score output by the gating network is temperature-scaled and / or normalized to adjust the difference in affinity scores between different expert networks, thereby improving the discrimination capability and stability of the distortion-aware expert routing module.

4. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 3, characterized in that, The distortion-aware expert selection routing module employs a capacity-limited expert claiming strategy for route allocation. Under this strategy, for each expert network, only the K distortion feature tokens with the highest affinity scores in the corresponding dimension of the expert network are allowed to claim from all distortion feature tokens. Furthermore, the number of distortion feature tokens claimed by the expert network does not exceed a preset capacity limit, thereby limiting the computational load of a single expert network during the routing phase. Here, K is the preset capacity parameter of the expert network.

5. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 4, characterized in that, When multiple expert networks attempt to claim the same distortion feature token, the distortion-aware expert routing module performs conflict resolution based on the following rules: Compare the affinity scores of the distorted feature tokens in each expert network that attempted to claim them; Distortion feature tokens are allocated to the expert network with the highest affinity score and whose current number of distortion feature tokens claimed has not reached the preset capacity limit. The successfully assigned distortion feature token will no longer participate in the claiming process of other expert networks.

6. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 5, characterized in that, The distortion feature token corresponding one-to-one with the image token feature is determined to be unclaimed by any expert network when any of the following conditions are met: The distorted feature token was not included in the claim set of the top K distorted feature tokens sorted from high to low based on affinity scores in all expert networks. The distorted feature tokens are entered into a claim set consisting of the top K distorted feature tokens sorted from high to low based on affinity scores by at least one expert network, but the at least one expert network has reached its capacity limit; If the distortion feature token is not claimed by any expert network, the distortion perception expert selection routing module will allocate the distortion feature token to an alternate processing path for feature processing to ensure that all distortion feature tokens are covered.

7. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 6, characterized in that, During training, the distortion-aware expert selection routing module introduces load balancing constraints. These constraints limit the distribution of the number of distortion feature tokens received by different expert networks during the training period to meet a preset deviation range, thereby preventing expert networks from being left unassigned processing tasks for extended periods. The distortion-aware expert routing module achieves a controllable trade-off between super-resolution reconstruction quality and computational resource consumption by adjusting the preset capacity parameters of the expert network.

8. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 1, characterized in that, The parameter-sharing hybrid expert network module includes multiple parallelly configured expert sub-networks, each of which performs nonlinear feature mapping operations on the assigned image token features; wherein... Each of the expert subnetworks includes at least two fully connected mapping layers, and a nonlinear activation function is set between adjacent fully connected mapping layers to perform nonlinear feature transformation and detail generation operations on the input image token features; The multiple expert subnetworks share the same backbone network structure and corresponding backbone parameters across different spectral bands.

9. The multispectral remote sensing image super-resolution method with distortion-aware routing and parameter sharing mechanism according to claim 8, characterized in that, In at least some of the fully connected mapping layers of the expert subnetwork, a low-rank adapter structure is introduced to achieve band-differentiated adjustment; the low-rank adapter structure includes a set of trainable weight matrices, which perform band-level incremental adjustment on the shared backbone parameters through low-rank matrix decomposition to achieve cross-spectral parameter sharing and band-specific feature adaptation.

10. A multispectral remote sensing image super-resolution system with distortion-aware routing and parameter sharing mechanism, characterized in that, include: The shallow feature extraction module is used to encode the features of the input low-resolution multispectral remote sensing image to obtain image token features corresponding to spatial location. The distortion-aware feature extractor module is used to extract distortion cue features reflecting high-frequency missingness, edge degradation and noise level in local areas from the image token features in the frequency domain and spatial domain, respectively, and to explicitly encode the distortion cue features into distortion feature tokens that correspond one-to-one with the image token features. The distortion-aware expert selection routing module is used to receive the distortion feature token and the image token feature corresponding to the distortion feature token, calculate the affinity relationship between the image token feature and multiple expert networks through a gating mechanism, and perform expert selection routing under capacity constraints. The capacity constraints are used to limit the number of distortion feature tokens assigned to each expert network to be processed to not exceed a preset upper limit, so that image token features with different distortion features are adaptively assigned to different expert networks for processing. A parameter-sharing hybrid expert network module is used to perform nonlinear feature reconstruction on image token features assigned by the distortion-aware expert selection routing module, wherein multiple expert networks share expert backbone network parameters across different spectral bands and introduce band-level lightweight parameter adaptation for different spectral bands. An aggregation module is used to perform weighted aggregation of image token feature outputs from one or more expert networks. The weights of the aggregation are obtained by normalizing the affinity scores output by the gating mechanism, and are fused with the image token features obtained by the shallow feature extraction module through residual methods to form a layer-by-layer enhanced high-resolution feature representation; and The reconstruction module is used to convert the layer-by-layer enhanced high-resolution feature representation into a target high-resolution multispectral remote sensing image.

Images